openmp/libomptarget/deviceRTLs/nvptx/test/parallel/num_threads.c

   1 // RUN: %compile-run-and-check
   2
   3 #include <stdio.h>
   4 #include <omp.h>
   5
   6 const int WarpSize = 32;
   7 const int NumThreads1 = 1 * WarpSize;
   8 const int NumThreads2 = 2 * WarpSize;
   9 const int NumThreads3 = 3 * WarpSize;
  10 const int MaxThreads = 1024;
  11
  12 int main(int argc, char *argv[]) {
  13   int check1[MaxThreads];
  14   int check2[MaxThreads];
  15   int check3[MaxThreads];
  16   int check4[MaxThreads];
  17   for (int i = 0; i < MaxThreads; i++) {
  18     check1[i] = check2[i] = check3[i] = check4[i] = 0;
  19   }
  20
  21   int maxThreads1 = -1;
  22   int maxThreads2 = -1;
  23   int maxThreads3 = -1;
  24
  25   #pragma omp target map(check1[:], check2[:], check3[:], check4[:]) \
  26                      map(maxThreads1, maxThreads2, maxThreads3)
  27   {
  28     #pragma omp parallel num_threads(NumThreads1)
  29     {
  30       check1[omp_get_thread_num()] += omp_get_num_threads();
  31     }
  32
  33     // API method to set number of threads in parallel regions without
  34     // num_threads() clause.
  35     omp_set_num_threads(NumThreads2);
  36     maxThreads1 = omp_get_max_threads();
  37     #pragma omp parallel
  38     {
  39       check2[omp_get_thread_num()] += omp_get_num_threads();
  40     }
  41
  42     maxThreads2 = omp_get_max_threads();
  43
  44     // num_threads() clause should override nthreads-var ICV.
  45     #pragma omp parallel num_threads(NumThreads3)
  46     {
  47       check3[omp_get_thread_num()] += omp_get_num_threads();
  48     }
  49
  50     maxThreads3 = omp_get_max_threads();
  51
  52     // Effect from omp_set_num_threads() should still be visible.
  53     #pragma omp parallel
  54     {
  55       check4[omp_get_thread_num()] += omp_get_num_threads();
  56     }
  57   }
  58
  59   // CHECK: maxThreads1 = 64
  60   printf("maxThreads1 = %d\n", maxThreads1);
  61   // CHECK: maxThreads2 = 64
  62   printf("maxThreads2 = %d\n", maxThreads2);
  63   // CHECK: maxThreads3 = 64
  64   printf("maxThreads3 = %d\n", maxThreads3);
  65
  66   // CHECK-NOT: invalid
  67   for (int i = 0; i < MaxThreads; i++) {
  68     if (i < NumThreads1) {
  69       if (check1[i] != NumThreads1) {
  70         printf("invalid: check1[%d] should be %d, is %d\n", i, NumThreads1, check1[i]);
  71       }
  72     } else if (check1[i] != 0) {
  73       printf("invalid: check1[%d] should be 0, is %d\n", i, check1[i]);
  74     }
  75
  76     if (i < NumThreads2) {
  77       if (check2[i] != NumThreads2) {
  78         printf("invalid: check2[%d] should be %d, is %d\n", i, NumThreads2, check2[i]);
  79       }
  80     } else if (check2[i] != 0) {
  81       printf("invalid: check2[%d] should be 0, is %d\n", i, check2[i]);
  82     }
  83
  84     if (i < NumThreads3) {
  85       if (check3[i] != NumThreads3) {
  86         printf("invalid: check3[%d] should be %d, is %d\n", i, NumThreads3, check3[i]);
  87       }
  88     } else if (check3[i] != 0) {
  89       printf("invalid: check3[%d] should be 0, is %d\n", i, check3[i]);
  90     }
  91
  92     if (i < NumThreads2) {
  93       if (check4[i] != NumThreads2) {
  94         printf("invalid: check4[%d] should be %d, is %d\n", i, NumThreads2, check4[i]);
  95       }
  96     } else if (check4[i] != 0) {
  97       printf("invalid: check4[%d] should be 0, is %d\n", i, check4[i]);
  98     }
  99   }
 100
 101   return 0;
 102 }